Publications Repository - Helmholtz-Zentrum Dresden-Rossendorf

Aim: Gaussian filters are commonly used to improve signal to noise ratio (SNR) of PET images but reduce spatial resolution, thus increasing partial volume effects. Edge preserving filters can alleviate this problem. Especially, the bilateral filter (BF) has shown to have very good performance but requires manual tuning of its two parameters (σS, σI) for optimal results. This is time consuming and hampers clinical use. In this work we, therefore, investigated an automatic method for parameter optimization.

Methods: PET-data from 69 patients were included: 18F-FDG(N=33), 18F-LDOPA(N=25), 68Ga-DOTATATE(N=11). All scans were performed with respiratory gating (8 gates), resulting in 552 low SNR PET volumes. Four 3D ROIs were placed in each volume: one within the liver (to assess noise) and three in areas with elevated focal uptake and various SNR levels. Optimal parameters were determined by a grid search in the (σS,σI) plane aiming at parameters that simultaneously leave SUV_max of the focal uptake mostly unaltered while yielding a noise level comparable to that seen in the sum over all gates. BF with the optimized parameters was then applied and images were visually inspected and analyzed regarding ΔSUV_max and ΔNoise differences (BF vs. unfiltered) in the respective ROIs.

Results: In 19/69 datasets our method failed (over-smoothed background or artifacts). For these images the parameters had to be manually tuned. Overall, optimal parameter values varied over a substantial range (mean±sd: σI=(1.4±1.5) SUV and σS=(5.5±1.7) mm) with σI exhibiting a pronounced tracer dependance. ΔSUV_max of the focal uptake ROIs across all datasets was small (-0.5±0.8) while substantial noise reduction was achieved by (-12.3±3.5) percentage points although detailed behaviour differed between tracers.

Conclusions: Our results demonstrate inter-individual and tracer-specific variability of optimal BF parameters and thus underline the need for careful parameter optimization. In 72% of all investigated cases our automated method was able to perform this optimization without any user intervention. More work is needed to further improve the success rate. However, already in its current form our method does notably reduce workload imposed on the user when considering BF for routine use.